Brief IA
›

Gemini 4 Hydrogen: Fast Execution, 1M Tokens, and Orchestrator

Research·Tom Levy·

Gemini 4 Hydrogen: Fast Execution, 1M Tokens, and Orchestrator

Gemini 4 Hydrogen: Fast Execution, 1M Tokens, and Orchestrator
Key Takeaways
1Gemini 4 Hydrogen enhances research with more visual responses and advanced query understanding
2Orchestrator enables dynamic switching between models and reduces quota usage
3Hydrogen shows speed gains and high scores on multiple benchmarks, particularly against Gemini 3.8 Flash
4The model is accessible through various Google services with a context of 1 million tokens and token-based pricing
Why it matters : Gemini 4 Hydrogen aims to make advanced AI faster and more accessible for developers and businesses, while integrating large-scale security protections.

Google places a new model at the heart of daily usage: Gemini 4 Hydrogen. The tool claims speed gains, multi-model orchestration, and extensive integration, from API to search, with a context of 1 million tokens and token-based pricing.

More Visual Search and Enhanced Protections Deployed at Scale

Google asserts that Gemini 4 Hydrogen improves search for all users globally, providing more useful and intuitive answers. Queries can be phrased in a natural thought flow, even if they contain multiple parts, difficult typos, or very specific requests, and the service would understand the intent. The responses are described as more adaptive, visual, and easy to navigate, at the expected speed. Thanks to its advanced reasoning capabilities, Hydrogen allows for an expanded range of queries and offers a broader array of relevant links than before. At the same time, Google specifies that it has added safety and security measures from Gemini 4, combining integrated learning and real-time surface protections to limit misalignments and abuses for the largest audience of Gemini 4 to date.

Benchmark Scores: Top Category Rankings and Quota Gains

In evaluations cited by Google, Orchestrator achieves success rates close to Carbon on Terminal-Bench 4.0 while consuming 24% less quota and completing tasks 19% faster. Hydrogen is ranked fifth on Agent Arena from arena.ai for multi-step tasks. It leads its category on HLE-Diamond with 58.8%, ranks first overall on Vals Tax Agent Bench with 52.3%, ties for first place on Vals Legal Research Bench with 55.8%, and claims the top category position on AutomationBench-AA.

Multi-Model Orchestration: Dynamic Switching and Parallel Agents

Hydrogen can be used alone or integrated into multi-model pipelines as an execution engine. Google introduces Orchestrator in Antigravity, a component that dynamically switches between models like Carbon and Hydrogen within the same session. For flows with parallel agents, Carbon can act as a coordinator while Hydrogen powers work and verification agents to build and test modifications simultaneously in isolated workspaces.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Announced Gains Against 3.8 Flash: 2.5x Speed and 57% on Terminal-Bench 4.1

Google presents Hydrogen as a major upgrade for migrating teams from Gemini 3.8 Flash. The model would complete tasks 2.5 times faster at its high default setting and shows 57% on Terminal-Bench 4.1 compared to 18.2% for 3.8 Flash. Reported claims would be reduced by more than five times, and the model adapts when tool calls fail or searches return no results. Google adds that Hydrogen would match or surpass competitors in its category on essential workflows.

Developers: Targeted Iteration, Parallel File Access, and Increased Speed

For coding and debugging repositories, Hydrogen is described as making targeted modifications without touching unrelated files and outperforming similar-level models on SWE-sweep. It natively groups file read and search calls in parallel to efficiently explore large repositories. According to partners in preliminary testing, it would operate 2 to 2.8 times faster than 3.8 Flash while reducing token consumption by 12% to 23% on complex graphs, SQL queries, and large data flows. For web development, it is capable of producing complete and operational layouts with fewer instructions in cases where execution speed is critical.

Multimodal and Real-World Use Cases: 1M in Input and Output, MKA 53.4% and ALE 71.6%

Hydrogen allows for native processing of up to 1 million tokens in input and up to 1 million in output to transform PDFs, spreadsheets, graphs, audio, or recorded videos into structured reports. Google specifies that it outperforms other models on the Multimodal Knowledge Agent benchmark with a score of 53.4% and achieves 71.6% on Agents’ Last Exam v1.1. In terms of applications, Geminisland relies on Hydrogen for real-time 3D coding, allowing users to add to a shared voxel island via a single prompt, with structures generated live and explorable in first-person view.

Access and Pricing: 1M Context and Token-Based Billing Announced

Hydrogen is made available in Google AI Studio, the Gemini API, Gemini for enterprises, the Gemini app, Google Search, and for paying subscribers of Google AI Pro and Ultra in Antigravity. The model is announced with a context of 1 million tokens and a fixed pricing of $1.50 per million tokens for input and $7.50 per million tokens for output. Google positions Hydrogen for rapid execution in agentic flows and daily coding, while Carbon and Argon are associated with deep reasoning. Teams can continue to rely on Carbon and Argon for demanding architectural decisions while mobilizing a fast and disciplined execution model for writing, testing, and verifying code.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.